Main

In your final repo, there should be an R markdown file that organizes all computational steps for evaluating your proposed Facial Expression Recognition framework.

This file is currently a template for running evaluation experiments. You should update it according to your codes but following precisely the same structure.

if(!require("EBImage")){
  install.packages("BiocManager")
  BiocManager::install("EBImage")
}
if(!require("R.matlab")){
  install.packages("R.matlab")
}
if(!require("readxl")){
  install.packages("readxl")
}

if(!require("dplyr")){
  install.packages("dplyr")
}
if(!require("readxl")){
  install.packages("readxl")
}

if(!require("ggplot2")){
  install.packages("ggplot2")
}

if(!require("caret")){
  install.packages("caret")
}

if(!require("glmnet")){
  install.packages("glmnet")
}

if(!require("WeightedROC")){
  install.packages("WeightedROC")
}
if(!require("DMwR")){
  install.packages("DMwR")
}
if(!require("mlr")){
  install.packages("mlr")
}
library(parallel)
library(parallelMap)
library(mlr)
library(R.matlab)
library(readxl)
library(dplyr)
library(EBImage)
library(ggplot2)
library(caret)
library(glmnet)
library(WeightedROC)
library(randomForest)
library(xgboost)
library(DMwR)
library(pROC)

Step 0 set work directories

set.seed(2020)
# setwd("~/Project3-FacialEmotionRecognition/doc")
# here replace it with your own path or manually set it in RStudio to where this rmd file is located. 
# use relative path for reproducibility

Provide directories for training images. Training images and Training fiducial points will be in different subfolders.

train_dir <- "../data/train_set/" # This will be modified for different data sets.
train_image_dir <- paste(train_dir, "images/", sep="")
train_pt_dir <- paste(train_dir,  "points/", sep="")
train_label_path <- paste(train_dir, "label.csv", sep="") 

Step 1: set up controls for evaluation experiments.

In this chunk, we have a set of controls for the evaluation experiments.

• (T/F) cross-validation on the training set
• (T/F) reweighting the samples for training set
• (number) K, the number of CV folds
• (T/F) process features for training set
• (T/F) run evaluation on an independent test set
• (T/F) process features for test set

run.cv <- FALSE # run cross-validation on the training set
sample.reweight <- TRUE # run sample reweighting in model training
K <- 5  # number of CV folds
run.feature.train <- TRUE # process features for training set
run.test <- TRUE # run evaluation on an independent test set
run.feature.test <- TRUE # process features for test set

Using cross-validation or independent test set evaluation, we compare the performance of models with different specifications. In this Starter Code, we tune parameter lambda (the amount of shrinkage) for logistic regression with LASSO penalty.

Step 2: import data and train-test split

#train-test split
info <- read.csv(train_label_path)
n <- nrow(info)
n_train <- round(n*(4/5), 0)
train_idx <- sample(info$Index, n_train, replace = F)
test_idx <- setdiff(info$Index, train_idx)

If you choose to extract features from images, such as using Gabor filter, R memory will exhaust all images are read together. The solution is to repeat reading a smaller batch(e.g 100) and process them.

n_files <- length(list.files(train_image_dir))

image_list <- list()
for(i in 1:100){
   image_list[[i]] <- readImage(paste0(train_image_dir, sprintf("%04d", i), ".jpg"))
}

Fiducial points are stored in matlab format. In this step, we read them and store them in a list.

#function to read fiducial points
#input: index
#output: matrix of fiducial points corresponding to the index
readMat.matrix <- function(index){
     return(round(readMat(paste0(train_pt_dir, sprintf("%04d", index), ".mat"))[[1]],0))
}

#load fiducial points
fiducial_pt_list <- lapply(1:n_files, readMat.matrix)
save(fiducial_pt_list, file="../output/fiducial_pt_list.RData")

Step 3: construct features and responses

• The follow plots show how pairwise distance between fiducial points can work as feature for facial emotion recognition.

  • In the first column, 78 fiducials points of each emotion are marked in order.
  • In the second column distributions of vertical distance between right pupil(1) and right brow peak(21) are shown in histograms. For example, the distance of an angry face tends to be shorter than that of a surprised face.
  • The third column is the distributions of vertical distances between right mouth corner(50) and the midpoint of the upper lip(52). For example, the distance of an happy face tends to be shorter than that of a sad face.

Figure1

feature.R should be the wrapper for all your feature engineering functions and options. The function feature( ) should have options that correspond to different scenarios for your project and produces an R object that contains features and responses that are required by all the models you are going to evaluate later.

• feature.R
• Input: list of images or fiducial point
• Output: an RData file that contains extracted features and corresponding responses

source("../lib/feature.R")
tm_feature_train <- NA
if(run.feature.train){
  tm_feature_train <- system.time(dat_train <- feature(fiducial_pt_list, train_idx))
  save(dat_train, file="../output/feature_train.RData")
}else{
  load(file="../output/feature_train.RData")
}

tm_feature_test <- NA
if(run.feature.test){
  tm_feature_test <- system.time(dat_test <- feature(fiducial_pt_list, test_idx))
  save(dat_test, file="../output/feature_test.RData")
}else{
  load(file="../output/feature_test.RData")
}

Step 4: Train a classification model with training features and responses

Call the train model and test model from library.

train.R and test.R should be wrappers for all your model training steps and your classification/prediction steps.

• train.R
  • Input: a data frame containing features and labels and a parameter list.
  • Output:a trained model
• test.R
  • Input: the fitted classification model using training data and processed features from testing images
  • Input: an R object that contains a trained classifier.
  • Output: training model specification

parameter tuning

1.First, you build the xgboost model using default parameters. You might be surprised to see that default parameters sometimes give impressive accuracy.

2.If you get a depressing model accuracy, do this: fix eta = 0.1, leave the rest of the parameters at default value, using xgb.cv function get best n_rounds. Now, build a model with these parameters and check the accuracy.

3.Otherwise, you can perform a grid search on rest of the parameters (max_depth, gamma, subsample, colsample_bytree etc) by fixing eta and nrounds. Note: If using gbtree, don’t introduce gamma until you see a significant difference in your train and test error.

4.Using the best parameters from grid search, tune the regularization parameters(alpha,lambda) if required.

5.At last, increase/decrease eta and follow the procedure. But remember, excessively lower eta values would allow the model to learn deep interactions in the data and in this process, it might capture noise. So be careful!

#preparing matrix 
feature_train = as.matrix(dat_train[, -6007])
label_train = as.integer(dat_train$label)-1 
#imbalanced data, needs smaller scale_pos_weight
sumwpos<-sum(label_train==1)
sumwneg<-sum(label_train==0)
dtrain<-xgb.DMatrix(feature_train,label=label_train)
run.cv_XGS.Step1<-FALSE
if(run.cv_XGS.Step1){
params <- list(booster = "gbtree", objective = "binary:logistic", eta=0.3, gamma=0, max_depth=5, min_child_weight=1, subsample=0.8, colsample_bytree=0.8,scale_pos_weight=sumwneg/sumwpos)
xgbcv <- xgb.cv( params = params, data = dtrain, nrounds = 200, nfold = 5, showsd = T, stratified = T, print_every_n = 10, early_stop_round = 20, maximize = F,eval_metric="auc",eval_metric="error")
 result<-xgbcv$evaluation_logt
 save(result, file="../output/res_cv_XGS.Step1.RData")
}else{
  load("../output/res_cv_XGS.Step1.RData")}

result_table<-rbind(result[which.min(result$test_error_mean),c("iter","test_error_mean","test_error_std","test_auc_mean","test_auc_std")],
result[which.max(result$test_auc_mean),c("iter","test_error_mean","test_error_std","test_auc_mean","test_auc_std")])

#best iter 160/199

#preprocessing the matrix for the test 
feature_test<-as.matrix(dat_test[,-6007])
label_test<- as.integer(dat_test$label)-1
dtest<-xgb.DMatrix(feature_test,label=label_test)


#first model trainig using
run.default.model=FALSE 
if(run.default.model){
xgb1<-xgb.train(params=params,
                data=dtrain,
                nrounds=160,
                watchlist=list(val=dtest,train=dtrain),
                print_every_n=10,
                early_stop_round=10,
                eval_metric="error",
                eval_metric="auc")

#model prediction
xgbpred.prob<- predict(xgb1,dtest)
xgbpred.label<- ifelse(xgbpred.prob>0.5,1,0)
levels(factor(xgbpred))
levels(factor(label_test))
library(caret)
confM<-confusionMatrix(factor(xgbpred.label),factor(label_test))
#accuracy 0.8433
#balanced accuracy 0.6886
auc<-roc(factor(label_test),xgbpred.prob)$auc
#auc:0.8489(high)
#  view variable importance plot 
mat<- xgb.importance(feature_names=colnames(feature_train),model=xgb1)
xgb.plot.importance(importance_matrix = mat[1:20])
save(confM,auc,mat,file="../output/res_cv_XGS.toy.RData")
}else{
  load("../output/res_cv_XGS.toy.RData")
}

#try to find the max depth and min child weight 
#source("../lib/cross_validation_XGS.R")
hyper_grid_XGB_1<- expand.grid(max_depth=3,#seq(3,10,by=2),
                              min_child_weight=1,#seq(1,6,by=2),
                              subsample=0.8, 
                              colsample_bytree=0.8
                               )
parallelStartSocket(cpus = detectCores())
dim(hyper_grid_XGB_1)
feature_train = as.matrix(dat_train[, -6007])
label_train = as.integer(dat_train$label) 
run.cv.XGS1=TRUE
if(run.cv.XGS1){
  res_cv_XGS1 <- matrix(0, nrow = nrow(hyper_grid_XGB_1), ncol = 4)
  for(i in 1:nrow(hyper_grid_XGB_1)){
    l<-list(max_depth=hyper_grid_XGB_1$max_depth[i],
            min_child_weight=hyper_grid_XGB_1$min_child_weight[i],
            subsample=hyper_grid_XGB_1$subsample[i],
            colsample_bytree=hyper_grid_XGB_1$colsample_bytree[i])
    xgbcv<-xgb.cv( params = l,
          data = dtrain, 
          nrounds = 160, 
          nfold = K, 
          gamma=0,
          showsd = T, 
          stratified = T, 
          early_stop_round = 10,
          print_every_n=80,
          eval_metric="auc",
          eval_metric="error",
          booster = "gbtree",
          objective = "binary:logistic", 
          scale_pos_weight=sumwneg/sumwpos,
          eta=0.3
          )
    result<-xgbcv$evaluation_log
    cv.error<-result$test_error_mean
    cv.AUC<-result$test_auc_mean
    res_cv_XGS1[i]<-c(mean(cv.error),sd(cv.error), mean(cv.AUC), sd(cv.AUC))
  save(res_cv_XGS1, file="../output/res_cv_XGS1.RData")
  }
}else{
  load("../output/res_cv_XGS1.RData")
}

#no need to tune gamma
hyper_grid_XGB_2<- expand.grid(max_depth=par_best_depth),
                              min_child_weight=par_best_child),
                              subsample=c(0.6,0.7,0.8,0.9), 
                              colsample_bytree=c(0.6,0.7,0.8,0.9)
                               )
parallelStartSocket(cpus = detectCores())
dim(hyper_grid_XGB_2)
feature_train = as.matrix(dat_train[, -6007])
label_train = as.integer(dat_train$label) 
run.cv.XGS2=TRUE
if(run.cv.XGS2){
  res_cv_XGS2 <- matrix(0, nrow = nrow(hyper_grid_XGB_1), ncol = 4)
  for(i in 1:nrow(hyper_grid_XGB_2)){
    l<-list(max_depth=hyper_grid_XGB_2$max_depth[i],
            min_child_weight=hyper_grid_XGB_2$min_child_weight[i],
            subsample=hyper_grid_XGB_2$subsample[i],
            colsample_bytree=hyper_grid_XGB_2$colsample_bytree[i])
    xgbcv<-xgb.cv( params = l,
          data = dtrain, 
          nrounds = 160, 
          nfold = K, 
          gamma=0,
          showsd = T, 
          stratified = T, 
          early_stop_round = 20, 
          eval_metric="auc",
          eval_metric="error",
          booster = "gbtree",
          objective = "binary:logistic", 
          scale_pos_weight=sumwneg/sumwpos,
          eta=0.3
          )
    result<-xgbcv$evaluation_log
    cv.error<-result$test_error_mean
    cv.AUC<-result$test_auc_mean
    res_cv_XGS2<-c(mean(cv.error),sd(cv.error), mean(cv.AUC), sd(cv.AUC))
  save(res_cv_XGS2, file="../output/res_cv_XGS2.RData")
  }
}else{
  load("../output/res_cv_XGS2.RData")
}

#tune the regularization parameters 

 xgb.cv( params = params, data = dtrain, nrounds = 200, nfold = 5, showsd = T, stratified = T, print_every_n = 10, early_stop_round = 20, maximize = F,eval_metric="auc",eval_metric="error")

traintask<-makeClassifTask(data=dat_train,target="label")
testtask<- makeClassifTask(data=dat_test,target="label")
cv.para.search=TRUE 
if(cv.para.search){
#create learner 
lrn_1 <- makeLearner("classif.xgboost",predict.type = "response")
lrn_1$par.vals <- list( objective="binary:logistic", eval_metric="error",eval_metric="auc", nrounds=200L, eta=0.3,scale_pos_weight=sumwneg/sumwpos)

#set parameter space 
params_1<- makeParamSet(
  makeIntegerParam("max_depth", lower=3L, upper=10L),
  makeNumericParam("min_child_weight", lower=1L, upper=10L),
  makeNumericParam("subsample",lower = 0.5,upper = 1),
  makeNumericParam("colsample_bytree",lower = 0.5,upper = 1)
)

#set resampling strategy 
rdesc_1<- makeResampleDesc("CV", stratify=T, iters=5L)
#search strategy 
ctrl_1<- makeTuneControlRandom(maxit=100L)#100 models
#set parallel backend 
parallelStartSocket(cpus = detectCores())
mytune_1 <- tuneParams(learner = lrn_1, task = traintask, resampling = rdesc_1, measures = acc, par.set = params_1, control = ctrl_1, show.info = T)
save(mytune_1, file="../output/res_cv_XGS.Step2.RData")
}else{
   load("../output/res_cv_XGS.Step2.RData")
}
mytune_1$y
#set hyperparameters 
lrn_tune_1 <- setHyperPars(lrn_1,par.vals = mytune_1$x)
#train model 
xgmodel_1 <- train(learner = lrn_tune_1,task = traintask)
#predict model 
xgpred_1.prob<- predict(xgmodel_1,testtask)
confusionMatrix(xgpred_1$data$response,xgpred_1$data$truth)# accracy=0.835, balanced accuracy=0.6544
auc_1<-roc(factor(label_test),xgbpred_1.prob)$auc #0.84 same as the default

#Result:Op. pars: max_depth=10; min_child_weight=9.93; subsample=0.846; colsample_bytree=0.627
#acc.test.mean=0.8441751, smaller than the default, why??????
#need auc and accuracy from confusion matrix 

# regulariztion tuning 
#high dimention using alpha to have a try 
#may be 
param_alpha<-

Summarize Running Time

Prediction performance matters, so does the running times for constructing features and for training the model, especially when the computation resource is limited.

cat("Time for constructing training features=", tm_feature_train[1], "s \n")
cat("Time for constructing testing features=", tm_feature_test[1], "s \n")
cat("Time for training model=", tm_train[1], "s \n") 
cat("Time for testing model=", tm_test[1], "s \n")

###Reference - Du, S., Tao, Y., & Martinez, A. M. (2014). Compound facial expressions of emotion. Proceedings of the National Academy of Sciences, 111(15), E1454-E1462.